Diagnose runtime and horizontal widths
The completed comparison is frozen. Its 16 time-limit failures cannot be retried, and its unused allowance cannot fund a follow-up.
The runtime work is complete for the current implementation: 24 native controls preserve all 348 bounds, matched short tests show a 9–11% improvement, and six full-history child-domain queries pass under the original time limit. Preserve those completed freezes; do not rerun them.
The saved endpoint diagnosis now reproduces all 18 fresh endpoints of the three complete groups. At the specified-field globe's eastward extremes, weighted training half-widths contribute 0.85–0.89 m and future remainder allowances 0.59 m. These are approximation-certificate terms, not measured aircraft errors. About 77% of that weighted training term comes from the generated 6.0-second training epoch. The future query is at 7.2 seconds.
Next:
- Trace how the eastward propagation allowance grows at those two query times in the fixed endpoint sectors. Keep source, force, rotation, coordinate and lever terms distinct; nested allowances cannot be added as independent errors. First qualify any added tracing against the unchanged native calculations.
- Use the trace to choose a tighter component-preserving propagation method. Preserve physical domains, training certificates, numerical centers, all coefficient columns and the original error/observation assumptions. Do not reduce a bound because its current value prevents qualification.
Every new trace or method comparison needs a separately costed finite freeze. Require native controls and a useful full-history screen before another complete six-group comparison. Do not repeat that comparison simply with more time or unchanged-width arithmetic.
Withheld aircraft fits remain gated by the complete width test, recording-scale qualification and the observational error model. ESA confirmation groups remain protected. Neither diagnostic authorizes an observed fit or a new scientific conclusion.